Convolutionalnetworksarepowerfulvisualmodelsthatyieldhierarchiesoffeatures.Weshowthatconvolutionalnetworksbythemselves,trainedend-to-end,pixelsto-pixels,exceedthestate-of-the-artinsemanticsegmentation.Ourkeyinsightistobuild“fullyconvolutional”networksthattakeinputofarbitrarysizeandproduce